Which LLM for long-context document processing in May 2026?

For 100K+ token document processing, use Gemini 3.1 Pro Preview (Google) as your default. It scores 57.2 on quality at $4.50/1M tokens and pushes 124 tok/s, which is the cleanest balance of comprehension and throughput available right now for jobs where the prompt itself is enormous. When you process long documents, most of your token spend is input, so a high-quality model at a moderate price beats a frontier model that charges twice as much for marginally better reasoning.

If your pipeline is throughput-bound rather than quality-bound, switch to Gemini 3.5 Flash at 227 tok/s and $3.38/1M tokens. It scores 55.3, two points below the Pro variant, but generates output 1.8x faster and costs 25% less. For summarization, extraction, and classification over large corpora, that trade is worth making. I would only reach past these two for documents where a single misread sentence carries real cost.

Why these picks and not the frontier models

The instinct is to grab the highest quality index. For long-context work that instinct is expensive and usually wrong.

Claude Opus 4.8 (Anthropic) leads on quality at 61.4, but it costs $10.00/1M tokens and runs at 63 tok/s. GPT-5.5 (OpenAI) sits at 60.2 quality and $11.25/1M. When you feed either of these a 150K-token contract, the input cost alone dwarfs the output, and you pay a premium for reasoning depth you rarely use on extraction and retrieval tasks.

The 4-point quality gap between Opus 4.8 and Gemini 3.1 Pro maps to a 2.2x price difference. On a document pipeline running millions of tokens per batch, that gap is the difference between a job that scales and one that gets rationed.

The throughput problem nobody prices in

Long-context jobs are latency-sensitive in an underappreciated way. A model running at 33 tok/s spends a long time producing output, and when you process documents in sequence that compounds.

This is why I rule out Kimi K2.6 (Kimi) for high-volume document work despite its attractive $1.42/1M price. At 33 tok/s it is the slowest model on this list. For a one-off analysis of a single filing, the price wins. For a queue of thousands of documents, the throughput tax erases the savings.

Qwen3.7 Max (Alibaba) is the throughput champion among high-quality models: 200 tok/s, 56.6 quality, $1.88/1M, and open source. If you can self-host and you want long-context processing without per-token billing, this is the model to deploy.

Top three for long-context work

Decision table

How to choose between Pro and Flash

Run both on a representative sample and measure parser failure rate on your actual output schema. If Flash produces structured output your downstream tooling accepts cleanly, take the 1.8x throughput and the lower price. If you see extraction errors on dense or ambiguous passages, the two quality points in Gemini 3.1 Pro pay for themselves in reduced manual review.

For most document pipelines the right answer is a tiered routing setup: Flash for the easy 80%, Pro for the documents that flagged low confidence. That keeps your blended cost near $3.50/1M while preserving accuracy where it matters.

Compare the candidates side by side in the LLM Selector, or browse the full field on Explore.